[Google Scholar] [OpenReview] [Find my email here]

I am an alumnus of Mila, University of Montreal, where I completed MSc studies in Computer Science with a concentration in AI.

During my MSc, I conducted research that was co-supervised by Prof. Aaron Courville and Prof. Pierre-Luc Bacon. I researched a subfield of AI called deep reinforcement learning, and contributed two works:

• Double Gumbel Q-Learning
  David Yu-Tung Hui, Aaron Courville, Pierre-Luc Bacon
  NeurIPS 2023 (Spotlight)
  We showed that function approximators introduce two heteroscedastic Gumbel noise sources to Q-learning. Modeling these noise sources almost doubled the performance of the popular SAC baseline.
  [.pdf] [Reviews] [Poster (.png)] [5-min talk] [1-hour seminar] [Code (GitHub)] [Errata]

• Stabilizing Q-Learning for Continuous Control
  David Yu-Tung Hui
  MSc Thesis, University of Montreal, 2022
  I derived a deep reinforcement learning algorithm from first principles. I derived loss functions from the principle of maximum entropy. And I derived a justification for regularizing the critic with LayerNorm using mathematical analysis inspired by the neural tangent kernel. Compared to baseline actor-critic algorithms, my algorithm did not diverge in high-dimensional continuous control.
  [.pdf] [Errata]

Before my MSc, I was a visiting researcher at Mila working with Prof. Yoshua Bengio. I designed algorithms that trained AI to follow instructions in a 2D gridworld game. I was a part of two projects:

• BabyAI 1.1
  David Yu-Tung Hui, Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Yoshua Bengio
  arXiv preprint, 2020
  We found three subtle changes to neural network architectures that almost tripled the sample efficiency of reinforcement learning for instruction following.
  [.pdf] [Code (GitHub)]

• Combating False Negatives in Adversarial Imitation Learning
  Konrad Zołna, Chitwan Saharia, Leonard Boussioux, David Yu-Tung Hui, Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Yoshua Bengio
  We presented three modifications to adversarial imitation learning that improved the sample efficiency of instruction following by at least an order of magnitude.
  IJCNN, 2021
  AAAI, 2020 (Student Abstract)
  [.pdf]

I also worked on two applications of deep learning during my time as a visiting researcher:

• Generating Multiscale Amorphous Molecular Structures Using Deep Learning: A Study in 2D
  Michael Kilgour, Nicolas Gastellu, David Yu-Tung Hui, Yoshua Bengio, Lena Simine
  The Journal of Physical Chemistry Letters, 2020
  We trained PixelCNN to generate images of amorphous materials.
  [.pdf]

• A Benchmark of Medical Out of Distribution Detection
  Tianshi Cao, Chin-Wei Huang, David Yu-Tung Hui, Joseph Paul Cohen
  UDL Workshop, ICML 2020 (Oral)
  We benchmarked deep learning algorithms on how well they detected medical images outside their training data.
  [.pdf] [10-min talk] [Code (GitHub)]

Before my time at Mila, I studied for a BA in Computer Science at Trinity College Cambridge and completed internships at Facebook and Tractable.

I also hold a DipABRSM in piano performance.

^*There are multiple ways to write my name. In Latin script, my surname is "Hui" and my firstname is "David Yu-Tung." In Traditional Chinese characters, my family name is "許" and my given name is "宇同." Most people call me "David." Others call me "宇同" or "Yu-Tung."